You’ve Got Mail: A Randomized Field Experiment on Tax Evasion

(1)

INFORMS is located in Maryland, USA

Management Science

Publication details, including instructions for authors and subscription information:

http://pubsonline.informs.org

You’ve Got Mail: A Randomized Field Experiment on Tax Evasion

Kristina M. Bott, Alexander W. Cappelen, Erik Ø. Sørensen, Bertil Tungodden

To cite this article:

Kristina M. Bott, Alexander W. Cappelen, Erik Ø. Sørensen, Bertil Tungodden (2019) You’ve Got Mail: A Randomized Field Experiment on Tax Evasion. Management Science

Published online in Articles in Advance 22 Oct 2019 . https://doi.org/10.1287/mnsc.2019.3390

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and- Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support of claims made of that product, publication, or service.

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.) and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to transform strategic visions and achieve better outcomes.

For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

(2)

You ’ ve Got Mail: A Randomized Field Experiment on Tax Evasion

Kristina M. Bott,âAlexander W. Cappelen,âErik Ø. Sørensen,âBertil Tungoddenâ

aCentre for Experimental Research on Fairness, Inequality and Rationality—The Choice Lab, Norwegian School of Economics, 5045 Bergen, Norway

Contact:kristina.bott@nhh.no, https://orcid.org/0000-0002-2491-8778(KMB);alexander.cappelen@nhh.no,

https://orcid.org/0000-0002-3489-7085(AWC);erik.sorensen@nhh.no, https://orcid.org/0000-0002-7155-4188(EØS);

bertil.tungodden@nhh.no, https://orcid.org/0000-0002-4182-8491(BT) Received:July 15, 2017

Revised:January 19, 2019 Accepted:April 15, 2019

Published Online in Articles in Advance:

October 22, 2019

Abstract. We report from a large-scale randomized ﬁeld experiment conducted on a unique sample of more than 15,000 taxpayers in Norway who were likely to have misreported their foreign income. By randomly manipulating a letter from the tax authorities, we cleanly identify that moral suasion and the perceived detection probability play a crucial role in shaping taxpayer behavior. The moral letter mainly works on the intensive margin, while the detection letter has a strong effect on the extensive margin. We further show that only the detection letter has long-term effects on tax compliance.

History:Accepted by Yan Chen, behavioral economics.

Open Access Statement: This work is licensed under a Creative Commons Attribution 4.0 International License. You are free to copy, distribute, transmit and adapt this work, but you must attribute this work as“Management Science. Copyright © 2019 The Author(s).https://doi.org/10.1287/mnsc.2019.3390, used under a Creative Commons Attribution License:https://creativecommons.org/licenses/by/4.0/.”

Funding:This work was supported by the Research Council of Norway through its Centres of Excellence Scheme [Centre for Experimental Research on Fairness, Inequality, and Rationality (FAIR) Project 26 26 75; Grant 236995] and the New Opportunities for Research Funding Agency Cooperation in Europe (NORFACE) Welfare State Future Program [ﬁle number 462-14-030] and administered by FAIR–The Choice Lab, NHH Norwegian School of Economics.

Supplemental Material:The online appendix is available athttps://doi.org/10.1287/mnsc.2019.3390.

Keywords: taxation•tax evasion•tax morale•ﬁeld experiment

1. Introduction

A key challenge in all modern societies is to limit tax evasion, which causes large losses in government revenues and creates significant unfairness in society. It has, for example, been argued that the loss of government revenue amounts to 500 billion USD annually in the United States, corresponding to the size of the government deficit, and 11 billion euros in Greece, corresponding to 30% of the government deficit (Cebula and Feige 2012, Artavanis et al. 2016, Cobham and Jansky 2018). Tax evasion is particularly difficult to handle when the tax administration has to rely on self- reported data, since taxpayers have an economic in- centive to underreport income (Allingham and Sandmo 1972, Slemrod and Yitzhaki2002, Sandmo2005). The classical approach to increasing tax compliance has therefore been to reduce the economic incentives for tax evasion by increasing the detection probability and penalties. In recent years, however, there has been a growing interest among policymakers and academics in understanding the extent to which moral motivation or, more broadly,“tax morale,”can play a role in increasing tax compliance in society, since it has the potential to ensure tax compliance even in situations where the capacity to control and audit taxpayers is limited (Slemrod2007, Luttmer and Singhal2014).

Tax morale is an umbrella term that captures non- pecuniary motivations for tax compliance (Luttmer and Singhal 2014), including fairness concerns, social norms, and reciprocal motives (Cullis and Lewis1997, Blumenthal et al. 2001, Torgler 2002, Wenzel 2004, Feld and Frey 2007, Kirchler 2007, Torgler 2007, Kirchler et al.2008). Many studies have pointed to tax morale as an explanation for why the level of tax compliance is often higher than what we would expect if people were motivated only by the fear of detection and penalties (Alm et al. 1992, Andreoni et al. 1998, Bobek and Hatﬁeld 2003), and several studies suggest that cross-country differences in tax compliance may partly reﬂect differences in tax morale across countries (Alm et al. 1995, Alm and Torgler 2006). Recent work has also provided neural evidence underscoring the importance of moral sentiments in the tax compliance context (Dulleck et al.2016).

To study the drivers of tax compliance, and in particular the role of moral motivation, we conducted a large-scaleﬁeld experiment together with the Nor- wegian Tax Administration (NTA) on a unique sample of more than 15,000 taxpayers. The sample consisted of taxpayers who were likely to have misreported their foreign income for the previous tax year, but who were not aware that the Norwegian tax authorities

1

(3)

had information about this misreporting. Informa- tion about foreign income is not included in the prepopulated tax return in Norway and the taxpayers therefore have to self-report this information. Historically, it has been difﬁcult for the tax authorities to verify the self- reported information because they have not had access to third-party reports from foreign countries, but over the last years this has changed due to international collaboration among tax authorities.

Theﬁeld intervention consisted of an information letter sent by the tax administration shortly before the taxpayers were due to submit their tax return for the previous year, where we randomly assigned taxpayers to receive different versions of a base letter or to a control group that did not receive any letter. The base letter contained information about why and how to report foreign income, and the effect of this letter sheds light on whether the underreporting was driven by a lack of information about tax-reporting procedures. The main aim of this study, however, was to identify the causal effects of introducing moral suasion and increasing the perceived detection probability of tax evasion, and we thus manipulated the base letter along each of these two dimensions in additional treatments. We study two versions of moral suasion—a fairness argument and a societal beneﬁts argument—

for correctly reporting foreign income. We assume that the main role of these letters is to make the moral argument for tax compliance salient, in the same way as Chetty et al. (2009) found that posting sales tax on price tags served to make the post-tax price more salient to consumers. To investigate the importance of the detection probability for tax evasion, we added information to the base letter that we believed would make the taxpayers increase their subjective probability of being audited.

We study the effect of our treatment manipulations on the self-reported foreign income both in the following tax return (the follow-up year) and one year later (long term).

We also investigate whether the effects are largely on the extensive margin (how many taxpayers self-report any foreign income) or on the intensive margin (how much foreign income is self-reported by taxpayers who would have self-reported some foreign income in the absence of the treatment), and we provide evidence on the un- derlying mechanisms of the treatment effects.

Our main result is that moral suasion has a large and signiﬁcant effect on self-reported foreign income.

For the follow-up year, the average self-reported foreign income by the taxpayers who received one of the moral letters was almost double the amount self-reported by those who received the base letter.

We alsoﬁnd a large effect of the detection letter, but the moral letters and the detection letter affect different margins of the taxpayer behavior: the detection letter has a large effect on the extensive margin, whereas the moral letters only have a large effect on

the intensive margin. Further, we show that the base letter itself has some effect on self-reported foreign income, but, overall, our study suggests that the underreporting is not primarily driven by a lack of knowledge about how to report foreign income.

Our ﬁndings are robust across different subgroups (age, gender, Norwegian citizenship, socioeconomic status). For all subgroups, we observe that the moral letters and the detection letter increase the level of self- reported foreign income for the follow-up year. Finally, we study the long-term effects of the intervention, where the main insight is that the detection letter has a large effect on the extensive margin even one year after the taxpayer received the letter, whereas there are no statistically signiﬁcant long-term effects of the moral letters.

These long-termﬁndings suggest that the moral letters mainly worked through making the moral arguments salient when the taxpayer received the letter, whereas the detection letter caused the taxpayers to permanently update their beliefs about the detection probability.

Our paper contributes to the growing literature using ﬁeld interventions to study tax compliance (Coleman1996, Blumenthal et al.2001, Slemrod et al.

2001, Torgler2004, Coleman2007, Kleven et al.2011, Ariel 2012, Del Carpio 2013, Fellner et al. 2013, Hallsworth 2014, Castro and Scartascini 2015, Pomeranz2015, Dwenger et al.2016, Hallsworth et al.

2017). Evidence from thefield experimental literature on the drivers of tax evasion has been mixed and, in particular, most of the studies have not been able to document that moral suasion may play an important role in reducing tax evasion. The pioneering paper by Blumenthal et al. (2001) uses a randomizedfield intervention in Minnesota to investigate the effect of moral appeals. In different letters sent to taxpayers from the tax authorities, they include a message highlighting the social benefits of servicesfinanced by tax revenues and a message indicating that the majority of citizens comply with tax laws. Theyfind no effect of these letters on tax compliance. The paper by Castro and Scartascini (2015) reports from a large field experiment in Argentina on payment of property taxes. Theyfind no average effects on tax compliance of two messages aimed at affecting beliefs about other taxpayers’ behavior and beliefs about the use of re- sources by the government. Similarly, Dwenger et al.

(2016), in the context of German church taxes, and Fellner et al. (2013), in the context of TV licenses fees in Austria, find no effect of appealing to morals or providing information about other’s compliance. How- ever, there are also interesting studies of tax compliance in the laboratory, which provide evidence of the potential role of moral suasion and taxpayer agency in reducing tax evasion (Dal B ó and Dal B ó 2009, Lamberton et al.2018) and of people overestimating the probability of detection (Alm et al.1992).

(4)

We believe our study has several strengths that may contribute to explaining why we can cleanly identify strong effects of both moral suasion and information that should increase the perceived detection probability. First, we consider a sample and a situation where there is no third-party reporting and all taxpayers have an opportunity to evade taxes, whereas some previous studies have suffered from a significant part of the sample being restricted by third-party reporting (Kleven et al. 2011). Second, we carefully timed the distribution of the letters such that the taxpayers received them close to the deadline for submitting their tax returns, whereas some previous studies had a significant lag between thefield intervention and the moment of decision making (Blumenthal et al.

2001, Slemrod et al. 2001). Third, our experimental design allows for a clean test of whether moral suasion or information that should increase the perceived detection probability drive the change in taxpayers’

behavior. We compare the effect of the moral letters and the detection letter to the effect of a base letter that only differs in the relevant dimensions, and we provide additional survey evidence showing that the letters worked as intended. Overall, we therefore believe that our study provides novel, clean, and robust evidence of how both moral suasion and information about detection probability may contribute to reducing tax evasion. Our study also contributes to the literature by showing that moral suasion and detection probability affect different margins of taxpayer behavior, and by providing, to our knowledge, theﬁrst set of results on the long-term effects of a ﬁeld intervention aiming to reduce tax evasion.

Taken together, our results shed light on the funda- mental question of why people pay taxes. In particular, the results contribute to a better understanding of how tax morale affects tax compliance, and suggest that it is possible to achieve a signiﬁcant increase in tax compliance by making moral appeals. However, our long- termﬁndings show that it is not trivial to strengthen tax morale in society. Moral appeals may only have short- term effects, and more comprehensive measures are therefore most likely needed to establish and sustain a tax morale that can foster compliance.

The structure of the paper is as follows: Section2pre- sents the setting for theﬁeld experiment and the sample, Section3provides details of the experimental design.

In Section4, we provide a simple theoretical framework to guide our analysis, and we outline the empirical strategy in Section5. Section6 discusses the results, and Section7provides some concluding remarks.

2. Background and Sample

In this section, we provide an overview of how taxes are reported in Norway. We then describe the sample of taxpayers in our study.

2.1. Tax Reporting in Norway

Every year in April, the NTA sends a prepopulated tax return for the previousﬁscal year to all Norwegian tax residents. The prepopulated tax return constitutes a preliminary tax statement and the taxpayer is re- quired to add any missing information and correct potential mistakes before the end of April. Taxpayers who believe the information in the prepopulated tax return to be correct and complete need not take any action.

Whenfiling their taxes, taxpayers are reminded to declare all income, both domestic and foreign, earned in the previous fiscal year. The domestic income is typically for the most part included in the prepopulated tax return, based on third-party reporting in Norway, but information about foreign income must be self-reported by the tax subjects. Historically, it has been difficult for NTA to audit whether taxpayers correctly report foreign income because there has been limited exchange of information across national tax jurisdictions. Over the last few years, however, tax administrations in a number of countries have increasingly provided information about the income and wealth that tax residents of other countries earn or hold in their countries. As part of this development, NTA has in recent years received reports from other tax administrations about Norwegian tax residents’

income and wealth in the respective countries. These reports are referred to as automatic country reports from abroad (ACA, in Norwegian,Automatiske Kon- trolloppgaver Utland). The exchange of such reports is a result of bilateral negotiations between national tax authorities, and not all countries exchange this type of information with NTA.

In the period we studied, NTA received these reports with delay and thus could not include information about foreign income in the prepopulated tax return. However, NTA could compare the self- reported foreign income in the domestic tax returns with the foreign income recorded in the ACA-report at a later date, and this comparison forms the basis for our study.

2.2. The Sample

Forﬁscal year 2011, which is the baseline year of this study, NTA received ACA-reports for around 40,000 Norwegian tax residents. NTA estimated that 17,899 of these had self-reported between 2,000 Norwegian Krone (NOK) and 200,000 NOK (equivalent to ap- proximately 350–35,000 USD in 2011) less in foreign income than stated in the ACA-reports for the income year 2011, and this group was the point of departure of the present study. The taxpayers were not aware that NTA had received reports about their foreign income, and the tax authorities did not act on this information until after the taxpayers had

(5)

self-reported foreign income for 2012, which we refer to as the follow-up year in this study. A small subset of the group that self-reported incorrectly was randomly selected to be part of a practical policy experiment that tested the usefulness of a weblink providing further information about how to report foreign income. In addition, we had to exclude 137 individuals for whom NTA had incomplete ACA- reports. The remaining 15,708 individuals constitute the sample used in our analysis.

The ACA-reports suggest voluntary compliance from about 55% of the taxpayers in a situation where they have discretion to misreport. This level of voluntary compliance is in line with what Kleven et al.

(2011) and Engstr öm and Holmlund (2009) find for self-employed individuals in Denmark and Sweden, respectively, and Castro and Scartascini (2015) find for property owners in Argentina (with respect to property tax), which are groups that have substantial discretion over paying taxes.

Table 1 provides a comparison of the general population and the tax subjects with an ACA-report on foreign income at baseline (the two left columns) and a comparison of those who self-reported foreign income correctly and those who self-reported incorrectly (the two right columns), where tax subjects are classiﬁed as self-reporting correctly if they have misreported less than 2,000 NOK for the baseline year.

From the left part of Table1, we observe that compared with the general population, the tax subjects with foreign income are more likely to be non-Norwegian citizens, a large share of whom are from other Nordic countries. We also observe that the individuals with foreign income are slightly more likely to be male and self-employed and are on average a few years older than the general population. From the right part of Table1, we observe that those who self-reported correctly and

those who self-reported incorrectly are very similar with regard to background characteristics, with the exception of those who self-reported incorrectly on average being older than those who self-reported correctly.

Table 2 provides descriptive statistics on income, wealth, and misreporting at baseline. From panels A and B, we observe that the tax subjects with ACA- reports have more income and wealth than the general population. Those who self-reported incorrectly have lower income than those who self-reported correctly, whereas the two groups have the same level of wealth. Those who self-reported incorrectly have the same level of income as the general population in Norway, but substantially more wealth.

From panels C and D, we observe that average foreign income in the ACA-reports is 44,902 NOK and the mean amount misreported is 8,866 NOK.

Those who self-reported incorrectly have substantially more foreign income in the ACA-reports than those who self-reported correctly (56,280 NOK versus 36,852 NOK) and self-reported only 51% of it to the tax authorities. Those who self-reported correctly reported 5,049 NOK more than what is stated in the reports from the tax administrations in other countries, which might reﬂect that the ACA-reports do not capture all foreign income for the tax subjects (both because the ACA-report from each country is likely to be incomplete and because NTA does not receive information from all tax authorities across the world).

3. Experimental Design

The basic structure of the experimental design is that all individuals in our sample received a prepopulated tax return for the follow-up year in week 14 of 2013, and individuals in the treatment arms then received a letter from the Norwegian tax authorities in week Table 1. Background Characteristics

Samples With ACA-report

Characteristic With ACA-report General population Incorrect Correct

Share Norwegian citizen 0.522 0.836 0.550 0.503

Share citizen of other Nordic country 0.456 0.039 0.431 0.474

Share female 0.445 0.502 0.455 0.437

Mean age 53.4 49.8 58.4 49.9

Share older than 60 years old 0.429 0.289 0.566 0.332

Share self-employed 0.117 0.084 0.095 0.133

Share high SES 0.487 0.353 0.452 0.512

Observations 37,897 215,956 15,708 22,189

Notes. The table reports background characteristics at baseline. “With ACA-report” refers to the individuals with an ACA-report. “General population” refers to a random sample (5%) of the population in the Norwegian tax records that do not have an ACA-report. “Incorrect” denotes the set of individuals with an ACA-report who have self-reported foreign income incorrectly, and

“Correct”denotes the set of individuals with an ACA-report who have self-reported foreign income correctly.“High SES”is deﬁned as the taxpayer being in the upper 25% of the income and wealth distribution at baseline. SES, socioeconomic status.

(6)

15 about how to report foreign income in the tax return. The experiment has an intention-to-treat design, since we do not know how many of the taxpayers actually read the letter. The likelihood of reading the letter, however, should not differ across treatments, since there were no treatment differences in the design of the envelopes. At our request, NTA conducted an independent survey to test whether tax subjects in general read letters from the tax authorities. The results from this survey suggest that about 60%–70% of taxpayers read the letters they receive from NTA.

The individuals could make changes to the prepopulated tax return, including self-reporting of foreign income, until the deadline in week 18. The individuals were randomly allocated into a control group (No letter) or one of three treatment arms (Basetreatment,Moral treatments, orDetectiontreatment). The individuals in the control group did not receive any letter from the tax

authorities, whereas the individuals in the different treatment arms received different versions of the base letter. In the online appendix, we provide translated versions of all the letters (see Figures B1–B4). Our main interest is whether receiving such a letter increased the self-reported foreign income for the follow-up year.

We also have long-term data that allow us to study whether the letter intervention affected the self- reported foreign income one year later.

All the letters had a link to the NTA website and the phone number of a call center in the tax administration established for the purpose of this study. To standardize the answers to the taxpayers interacting with the call center, NTA provided phone operators with a script of potential questions and answers. The phone operators were not aware that the call center was facilitating a ﬁeld experiment; they only knew that the authorities had sent out different letters to different individuals. In Table3, we provide an overview of the activity at the call center. Five percent of the individuals receiving a letter approached the call center, mainly asking questions about why they had received the letter and how to report foreign income.

Signiﬁcantly more individuals used the call center in theDetectiontreatment than in the Base treatment (13.1%

versus 3.6%,p<0.001), whereas we see no difference between theMoraltreatments and the Base treatment (3.7% versus 3.6%,p0.710).

3.1. BaseTreatment

Since any letter from the tax authorities may cause a change in behavior for a number of reasons (fear of detection, moral salience, or better knowledge about how to proceed with reporting), we included a treatment where the tax residents received a letter that only contained general information about how to self-report foreign income (Basetreatment).

The letter consists of three paragraphs, the first explaining why the reader is receiving this letter. It refers to the fact that the Norwegian economy is becoming more international, with an increasing number of taxpayers having income from abroad, and states that NTA would like to inform taxpayers about how this type of income is taxed and how it should be reported. The second paragraph states that all Norwegian tax residents are liable to pay taxes to Norway on all income and assets, even on foreign income and foreign assets, unless otherwise specified in Norway’s tax treaties with other countries. Thefinal paragraph describes how to proceed after receiving the prepopulated tax return, and adds a weblink providing further information about how tofile the Norwegian tax return. We had two different versions of the base letter, one using active language, thus, addressing the reader as “you,” and another using passive language (Bryan et al. 2012). We do not find an economically or statistically significant Table 2. Income, Wealth, and Misreporting (Baseline)

Samples With ACA-report

With ACA-report General population Incorrect Correct Panel A: Taxable income

Mean 360,628 272,616 299,838 403,619

Q25 119,834 110,447 97,199 147,551

Q50 234,809 215,354 182,845 274,685

Q75 419,110 345,076 347,295 458,413

Panel B: Taxable wealth

Mean 1,330,938 462,820 1,530,805 1,189,590

Q25 0 0 0 0

Q50 43,248 63 58,366 35,277

Q75 609,583 325,706 651,296 577,269

Panel C: ACA-reports of foreign income

Mean 44,902 56,280 36,852

Q25 519 7,509 105

Q50 6,560 18,987 868

Q75 29,073 48,670 12,284

Panel D: Estimate of misreporting

Mean 8,866 28,533 −5,049

Q25 16 4,187 3

Q50 985 14,209 154

Q75 13,556 36,732 948

Notes. Panels A and B refer to (taxable) income and wealth at baseline. Panel C shows total foreign income in the ACA-reports, and panel D shows misreported foreign income (calculated by the difference between the foreign income in the ACA-reports and self- reported foreign income). The termQxrefers to thex-percentile in the relevant group. “With ACA-report” refers to individuals with an ACA-report.“General population”refers to a random sample (5%) of the population in the Norwegian tax records that do not have an ACA-report.“Incorrect”denotes the set of individuals with an ACA- report who have self-reported foreign income incorrectly, and

“Correct”denotes the set of individuals with an ACA-report who have self-reported foreign income correctly.

(7)

difference between these two versions of the base letter on the amount self-reported (p0.775) or on the share of individuals self-reporting a positive amount (p0.884), and thus we pool them in the analysis.

TheBasetreatment allows us to study whether lack of information about how to report foreign income is a main driver of the observed underreporting of foreign income. It is, however, important to note that the treatment difference between theBasetreatment and theNo lettergroup provides an upper bound estimate of the role of information, since the base letter may also trigger other mechanisms among taxpayers.

3.2. MoralandDetection Treatments

To identify as cleanly as possible the causal effects of introducing moral suasion and of increasing the detection probability, we manipulated the base letter along each of these two dimensions in additional treatments.

The treatment manipulations only introduced minor changes in theﬁrst paragraph of the letter; the rest of the letter was identical to the base letter.

We studied two types of moral appeals. In the Fairness treatment, the letter introduced a fairness argument for reporting foreign income correctly by including a sentence that reminded taxpayers that most Norwegians report the income earned in Nor- way correctly. Speciﬁcally, the following two sen- tences were added to the end of theﬁrst paragraph:

“The great majority report information about their income and assets in Norway correctly and completely. In order to treat all taxpayers fairly, it is therefore important that foreign income and foreign

assets are reported in the same manner.” The statement in this letter is correct. Most income earned in Norway is third-party reported to the tax authorities, and thus Norwegians have little discretion to evade taxes. An estimate of the share of individuals in Norway having all their income third-party reported is provided in Foss et al. (2015), who analyze a rep- resentative sample of 4,000 tax returns. They show that about 70% of Norwegians did not make adjust- ments to the preﬁlled tax return in 2010 and establish that there are very few errors in these tax returns, suggesting that these taxpayers only have third-party reported income. The numbers are in line with what is found for Denmark in Kleven et al. (2011), where 62%

did not self-report any additional income (p. 665, panel B). The difference between the estimates of Foss et al. (2015) and Kleven et al. (2011) is consistent with technological progress making third-party reporting increasingly common.

In the Societal benefits treatment, the letter introduced a societal benefits argument for reporting foreign income correctly by including a sentence that reminded taxpayers about the benefits to society resulting from taxation:“Your tax payment contributes to the funding of publicly financed services in education, health and other important sectors of society.”This sentence may trigger a reciprocity motive for tax compliance, where individuals become more willing to pay taxes because they recognize this as an exchange for benefits that the state provides (Fehr and G ächter 1998, Luttmer and Singhal 2014). In two additional treatments, we visualized the societal benefits from taxation by adding an attachment Table 3. Statistics from the Contact Logs

Statistic Total

Treatment group

Test of equality p-values Base Moral Detection Moral Detection Share of letter recipients in the logs 0.050 0.036 0.037 0.131 0.710 <0.001 Communication with letter recipient:

Question about why he/she receives letter 0.282 0.261 0.254 0.324 0.888 0.182

Negative reaction to letter 0.055 0.049 0.041 0.073 0.701 0.364

Question about rules of taxation 0.347 0.348 0.351 0.343 0.951 0.936 Arguing about legitimacy of rules 0.063 0.070 0.058 0.065 0.621 0.832 Question about how to report foreign income 0.721 0.741 0.758 0.668 0.709 0.126 Question about how foreign income is audited 0.019 0.007 0.014 0.031 0.540 0.127 Question about reporting income from before 2012 0.062 0.035 0.041 0.099 0.762 0.021 Characteristics of caller:

Mean age 67.3 70.4 68.3 64.4 0.105 <0.001

Share female 0.544 0.643 0.558 0.475 0.089 <0.001

Share Norwegian citizen 0.611 0.664 0.605 0.589 0.223 0.138

Observations 700 143 294 263

Notes. The table shows the share of taxpayers that contacted the call center, the share that asked each of the questions, and background characteristics. The table is based on 700 contacts with the call center in which callers identiﬁed which letter they had received.“Test of equality”reports a test of equality between the respective treatment and theBasetreatment. Thep-values are based on Pearsonχtests on binary outcomes (and t-tests for tests of mean age).

(8)

illustrating publicly ﬁnanced services in health, education, infrastructure, and research (see Figure B5 in the online appendix). In one treatment, the attachment was combined with the base letter; in another treatment it was combined with the societal beneﬁts letter.

In the Detection treatment, the letter aimed to increase the perceived detection probability of the tax subject. We replaced theﬁrst sentence in the base letter,

“The Norwegian economy is becoming more inter- nationalised, and an increasing number of Norwegian taxpayers receive income and have assets abroad,”

with,“The tax administration has received information that you have had income and/or assets abroad in previous years.”The basic idea behind this treatment manipulation was that providing information about the tax authorities’knowledge about the individual’s income-generating activities abroad in previous years would make the tax subjects update their subjective beliefs about the likelihood of being audited. How- ever, it is not obvious that the detection letter has a strong effect on tax compliance. The taxpayers may believe that the tax authorities only have information about some of their foreign income or they may not consider the letter to represent a credible threat about tax enforcement (Slemrod et al.2001).

Table 4 provides an overview of the different treatment arms in the experiment. The experimental design allows for the following two main comparisons to study the drivers of the misreporting of foreign income:

• The role of moral motivation: The comparison between theMoraltreatments and theBasetreatment identiﬁes the causal effect of moral suasion on taxpayer behavior.

• The role of the detection probability: The comparison between theDetectiontreatment and theBase treatment identiﬁes the causal effect of increasing perceived detection probability on taxpayer behavior.

When interpreting these treatment comparisons, we make two assumptions. First, we assume that the

moral letters only manipulate the moral dimension relative to the base letter. Second, we assume that the detection letter only manipulates the perceived detection probability relative to the base letter. We tested these assumptions in an independent survey, where, as shown in Table A1 in the online appendix, weﬁnd support for the letters working as intended.

Importantly, weﬁnd no evidence of the moral letters increasing the perceived detection probability among recipients of these letters.

4. Theoretical Framework

We here provide a simple model of taxpayer behavior to guide our analysis and the interpretation of the results, building on Cappelen et al. (2007) and Sandmo (2012). We do not model the behavior of the tax authorities and thus do not consider the possibility of strategic interaction between the taxpayers and the tax authorities (Graetz et al.1986).

Assume that the taxpayer has (only) foreign income yand self-reports the foreign income to ber. The tax on foreign income istand the penalty on misreported income, if detected, isτ. After tax, income isYy−tr if the misreporting is not detected and Zy−tr− τ(y−r) if it is detected. Let us assume that the taxpayer has the following expected utility function:

EU(r;·) pu(Z) + (1−p)u(Y) −s(T)β(y−r)₂ , (1) where p is the subjective probability of being detected, uis a von Neumann-Morgenstern utility function capturing the risk preferences over income,s(T) ∈ [0,1] is a parameter capturing the salience of the moral cost in treatment T, and β≥0 is the weight attached to the morals cost of misreporting when the moral cost is fully salient (s(T) 1). The model captures that peopleﬁnd it morally right to report their income correctly and that the moral cost is increasing in the deviation between their true income and their Table 4. Overview of Treatments

Treatment Description

No letter Did not receive a letter Base General information letter

Fairness Base letter + the following sentence added to theﬁrst paragraph:

“The great majority report information about their income and assets in Norway correctly and completely. In order to treat all taxpayers fairly, it is therefore important that foreign income and foreign assets are reported in the same manner.”

Societal beneﬁts Base letter + the following sentence added to theﬁrst paragraph:

“Your tax payment contributes to the funding of publiclyﬁnanced services in education, health and other important sectors of society.”

Two additional treatments included an attachment visualizing public servicesﬁnanced through taxes (without the base letter/in combination with the base letter).

Detection Base letter, but theﬁrst sentence (“The Norwegian economy is becoming more internationalized and an increasing number of Norwegian taxpayers receive income and have assets abroad”) is replaced by the following sentence:

“The tax administration has received information that you in previous years have had income and/or assets abroad.”

(9)

reported income. It follows straightforwardly that the interior solution for the taxpayer is given by

ry−Δu^s(r,t, τ,p)

2sβ , (2)

where Δu^s(r,t, τ,p) t(1−p)u(Y) − (τ−t)pu(Z). The second term in theﬁrst order condition captures the trade-off that determines the level of misreporting.

The nominator represents the marginal gain of misreporting at the interior solution, whereas the de- nominator shows the marginal cost. In the case where the moral cost is salient (s>0) and the taxpayer only cares about the moral cost (β→∞), the optimal choice is to self-report the foreign income correctly.

The Moral treatments and the Detectiontreatment aim to increase the weight attached to the moral cost of misreporting,sβ, and the subjective detection prob- abilityp, respectively, and thereby to increase the self- reported foreign incomer. We do not expect the moral letters to fundamentally change the strength of moral motivationβ, but they may make the moral argument more salient, causing an increase in s. TheDetection treatment provides new information to taxpayers, which should make them update their subjective beliefs about the likelihood of being detected, causing an increase inp.

The treatments may work both on the extensive and the intensive margin, but they do not necessarily work on the same margin. Let us provide an example.

Assume that there are two types of taxpayers:selﬁsh (β0) and moral (β>0). Further, assume that ev- eryone has a subjective probability of detectionp0 in theBase treatment, which implies that the selﬁsh taxpayers do not report any foreign income. Finally, assume that the moral taxpayers are at an interior solution and report some foreign income. In this case, it follows that theMoraltreatments only work at the

intensive margin: they increase the foreign income reported by the moral taxpayers by making the moral cost more salient, whereas the selﬁsh individuals will not be affected by an increase in the salience parameters. In contrast, theDetectiontreatment affects both the extensive and the intensive margin. It affects the extensive margin by increasing the subjective probability of detection and thereby making it more costly for the selﬁsh taxpayers not to report any foreign income, and it affects the intensive margin since moral taxpayers also care about the subjective probability of detection.

The main insights from this example apply to all situations where we have a separation of selﬁsh taxpayers not reporting any foreign income and moral taxpayers reporting some foreign income. In all such cases, only theDetection treatment works on the extensive margin, whereas the Moral treatments only work on the intensive margin. More generally, the treatment effects will depend on the tax parameters, the shape of the utility function, the subjective detection probability, and the salience and importance assigned to the moral cost of misreporting.

5. Data and Empirical Strategy

We here provide an overview of the data and the empirical strategy for the main analysis and the heterogeneity analysis.

5.1. Data

The analysis uses data from the administrative records of NTA. Self-reported foreign income in the 2011 tax return serves as a baseline variable, whereas the main outcome variable of interest is self-reported foreign income for 2012 (follow-up year) and 2013 (long term). Further, we use the ACA-reports received by NTA for 2011 and 2012 from 17 countries as Table 5. Baseline Characteristics by Treatment

Treatment n

Positive amount self-reported

for 2011

Amount self-reported

for 2011

Amount in ACA-reports

for 2011

Amount in ACA-reports

for 2012

Share female

Share above 60 years old

Share Norwegian

citizen

High socioeconomic

status

No letter 1,968 0.393 30,287 83,995 43,303 0.460 0.544 0.553 0.441

(0.011) (4,356) (33,390) (6,761) (0.011) (0.011) (0.011) (0.011)

Base 3,947 0.402 27,427 55,211 44,104 0.450 0.551 0.553 0.454

(0.008) (3,278) (3,935) (5,632) (0.006) (0.008) 0.008) (0.008)

Moral 7,821 0.405 28,040 51,971 38,097 0.459 0.561 0.548 0.454

(0.006) (2,784) (2,070) (2,637) (0.06) (0.006) (0.005) (0.006)

Detection 1,972 0.417 24,689 47,855 58,313 0.446 0.542 0.548 0.452

(0.011) (3,436) (2,777) (20,441) (0.011) (0.011) (0.011) (0.011)

Total 15,708 0.404 27,747 56,280 42,796 0.455 0.554 0.550 0.452

(0.004) (1,756) (4,434) (3,321) (0.004) (0.004) (0.004) (0.004)

Notes. The table reports the baseline characteristics used as controls in the main regression speciﬁcations, by treatment. Standard errors of the means are in parentheses.

(10)

background variables. According to our agreement with NTA, we are not allowed to list the countries providing ACA-reports to NTA. The administrative records also include data on age, gender, citizenship, income, and wealth.

Table5shows the background variables used in the analysis by treatment. We note that there are some differences across treatments in the amounts self- reported for 2011 and in the amounts in the ACA-reports for 2011 and 2012, but these differences are not statistically significant (one-way analysis of variance [ANOVA] test: amount reported 2011,p0.88; amount in ACA-reports 2011,p0.12; amount in ACA-reports 2012, p0.29.) For all other background variables, there is almost no variation across treatments.

5.2. Empirical Strategy

In the analysis, our main regression speciﬁcation is y_i,tα+∑

l∈L

βld_il+δyi,t0+γxi+εi, (3)

where y_i,t is self-reported foreign income for indi- vidualifor yeart. We letlindex a treatment in the set of treatmentsL, whered_ilis an indicator variable for whether individualiis in treatmentl;y_i,t₀ is the self- reported foreign income for the baseline yeart0, and xiis a vector of background variables (including age, gender, Norwegian citizenship, and a measure of socioeconomic status deﬁned by income and wealth). The estimated causal effect of treatment l relative to the reference treatment is then given by the estimatedβl

coefﬁcient.

We will start by reporting regressions where we pool all the treatment arms and deﬁne receiving a letter as the omitted category, which provides us with an estimate of the average causal effect of receiving any one of the letters from the tax authorities. We then report regressions where we estimate separately the average causal effects of the different letters relative to the Basetreatment, before we pool theMoral treatments to get an estimate of the average causal effect of theMoral treatments. We use the same approach to study the effect on the extensive margin, where the dependent variable is an indicator variable for self-reported foreign income for year t being strictly positive. For all speciﬁcations, we report estimates for regressions both with and without the background variables. We also report the effect on self-reported foreign income conditional on it being positive, but it is important to keep in mind that this conditional variable is determined both by the extensive margin and the intensive margin.

If we are willing to make the reasonable assumption of monotone treatment response with respect to a

Basetreatment (b), we can be more specific about the effect on the intensive margin (Manski 1997). Lety^l_i denote the self-reported foreign income of individual i in treatment l, and assume a monotone treatment response where y^l_i≥y^b_i for all individuals iin treatment l. This assumption, combined with random assign- ment to treatment, allows us to decompose the average treatment effect as ATE^lπ^bATE^l_b+π^Δ(l)ATE^l_Δ(l), where π^b is the share of taxpayers who reports a positive amount in theBasetreatment, and ATE^l_bis the average treatment effect on this group–the intensive margin–in treatmentl. The proportionπ^Δ(l)represents those that did not report anything in theBasetreatment, but that are induced to do so in treatmentl, and is the effect on the extensive margin:π^Δ(l)Pr(yi>0|Ll)−Pr(yi>0| Lb). For a treatment for which there is no effect on the extensive treatment,π^Δ(l)0, it follows that the effect on the intensive margin is given by ATE^l_b ATE^l/π^b. For a treatment with a positive extensive margin,π^Δ(l)>0, the effect on the intensive margin cannot be point estimated. However, since we assumey^l_i,t≥y^b_i,t, it follows that ATE^l_Δ(l)≥0, and we can bound the intensive margin: ATE^l_b∈ [0,ATE^l/π^b). Being more specific about the intensive margin with nonzero extensive margin effects would rely on parametrically modeling responses, and, as with other selection models, iden- tification would rely either on fully specifying the dis- tributions of all unobservable variables or on access to a variable that predicts reporting a positive amount without directly affecting the amount itself (Staub2014).

To study whether there are large differences in how the treatments affect the participants, we also conduct a heterogeneity analysis using the background variables age, gender, Norwegian citizenship, and socioeconomic status. In this analysis, we take the regression speciﬁcation where we have pooled the Moral treatments as the point of departure. For each background variable, we partition the set of participantsIintoG andI\G, withgias an indicator variable for whether individualiis a member ofG. To illustrate, if the relevant background variable is age, then we partition the set of participants into two subsets, those who are below and above 60 years. The indicator variable would then take the value one if the taxpayer is above 60 years. In each case, we interact the indicator variable with the treatment indicatord_il,

y_i,tα+βdil+θgid_il+λgi+δyi,t0+γxi+εi. (4) In this analysis, the estimation sample is the participants in the l treatment and in theBase treatment.

With this speciﬁcation, the estimated treatment effect of being in treatment lfor individuals in group Gis β+θ, whereas it isβfor individuals in groupI\G. The parameterθis the estimated difference in treatment

(11)

effect between the two groups, and it provides the basis for a statistical test of whether the estimated heterogeneity is statistically signiﬁcant.

6. Results

In this section, we examine how the treatments affected self-reporting for the follow-up year, before turning to the heterogeneity analysis. In theﬁnal part, we report on long-term effects of the intervention.

6.1. Main Analysis

From the ACA-reports, we know that 78% of the taxpayers in our sample had foreign income in the follow-up year, on average 42,796 NOK. We now study whether the different letters caused the taxpayers to self-report a larger share of their foreign income to tax authorities.

As shown in panel (a) of Figure1, for the follow-up year, the average self-reported foreign income of the taxpayers who received the base letter is larger than that of those who did not receive any letter. Strikingly, however, we observe that the self-reported foreign income of the taxpayers who received one of the moral letters or the detection letter is almost double the amount self-reported by those who received the base letter. In Table6, we report the corresponding regressions based on Equation (3). Columns (1) and (2) report estimates of the effect of not receiving a

letter, where all taxpayers who did receive a letter are pooled and serve as the reference category. We observe from column (1) that receiving a letter has a large and highly statistically signiﬁcant effect on self- reported foreign income: it more than doubles, from 8,155 NOK in theNo lettergroup to 17,030 NOK in the treatment groups combined. As shown in column (2), the ﬁnding is robust to the inclusion of a set of background variables, including self-reported foreign income for the baseline year, foreign income recorded in the ACA-report for the baseline year and the follow-up year, and personal and socioeconomic characteristics of the taxpayer (p<0.001, column (2)).

Finally, in panel (a) of Figure2, we show that there are no statistically signiﬁcant differences across treatments in the amount requested for deductions based on taxes paid abroad (p0.551), even though panel (b) shows that the letters cause a slight increase in the share of taxpayers requesting a deduction. Thus, we can state ourﬁrst main result.

Result 1. A letter from the tax authorities has a large and statistically signiﬁcant effect on the self-reported foreign income.

In columns (3) and (4), we report estimated treatment effects for each of the letters separately, where the base letter now serves as the reference treatment.

We observe that theNo lettertreatment group reports a lower foreign income than theBasetreatment group.

Figure 1. Self-Reported Foreign Income

Note. Theﬁgure shows, for the follow-up year, the average amount of self-reported foreign income in NOK (panel (a)) and the share of taxpayers who self-report a positive foreign income (panel (b)), by treatment.

(12)

The estimated effect of the base letter is positive, but not statistically signiﬁcant (p0.113, column (4)), which suggests that the underreporting is not primarily driven by a lack of information about how to report foreign income.

The estimates for the moral letters and the detection letter identify the causal effects of adding moral suasion and increasing the detection probability. We observe that the effects are large and highly signiﬁcant for all the three main treatments (Fairness:p0.010;

Societal beneﬁts: p0.034; Detection: p0.028, column (4)).¹Moral suasion, in terms of a fairness or societal beneﬁts argument, has a strong positive effect

on self-reported foreign income of almost the same magnitude as information that should increase the perceived detection probability. The argument presented in theFairness treatment explicitly appeals to the importance of treating all taxpayers fairly, and our results are thus in line with previous research in behavioral economics documenting that people are willing to sacriﬁce pecuniary gains to avoid large deviations from what they consider fair (Fehr and Schmidt1999, Bolton and Ockenfels2000, Andreoni and Miller2002, Charness and Rabin 2002, Cappelen et al. 2007).

However, the Fairnesstreatment also introduces a social comparison by emphasizing what the majority of Table 6. Treatment Effects on Self-Reported Foreign Income

Treatment (1) (2) (3) (4) (5) (6)

No letter −8,874.9*** −10,008.1*** −3,188.0* −4,339.1 −3,188.0* −4,339.6 (2,184.5) (2,767.4) (1,643.1) (2,734.8) (1,643.1) (2,735.2)

Fairness 15,158.5* 10,372.1**

(8,860.6) (4,041.9)

Societal beneﬁts 5,180.9** 6,345.5**

(2,596.8) (2,989.0)

Detection 9,199.6** 10,351.3** 9,199.6** 10,351.6**

(4,385.6) (4,702.7) (4,385.5) (4,702.8)

Moral 7,671.1** 7,350.5***

(3,010.7) (2,772.9)

Foreign income baseline 0.42** 0.42** 0.42**

(0.20) (0.20) (0.20)

ACA-report baseline 0.00089 0.00091 0.00091

(0.0035) (0.0035) (0.0035)

ACA-report follow-up 0.010 0.010 0.010

(0.0073) (0.0073) (0.0073)

Female 6,791.0* 6,782.8* 6,777.1*

(3,640.9) (3,634.1) (3,633.9)

Age>60 years 9,167.8* 9,148.3* 9,144.6*

(5,207.1) (5,209.7) (5,207.6)

Norwegian citizen 2,702.2 2,733.2 2,741.5

(3,294.8) (3,307.6) (3,306.1)

High SES 4,591.3 4,615.8 4,595.0

(5,275.6) (5,262.7) (5,274.6)

Constant 17,029.8*** −6,706.3 11,342.9*** −12,387.4 11,342.9*** −12,380.8 (1,752.6) (6,673.8) (999.6) (7,537.0) (999.6) (7,535.5) p-value fromF-test onMoral

treatments being equal: 0.27 0.32

Observations 15,708 15,708 15,708 15,708 15,708 15,708

R² 0.000 0.230 0.001 0.231 0.001 0.231

Notes. The table reports regressions based on Equation (3), where the dependent variable is the amount of foreign income self-reported for the follow-up year. In columns (1) and (2), the estimated effects are relative to the pooled sample of all treatment groups; in columns (3)–(6), the estimated effects are relative to theBase treatment. The indicator variablesNo letter,Fairness,Societal Beneﬁts, andDetectiontake the value one if the taxpayer is in the respective treatment. The indicator variableMoraltakes the value one if the taxpayer is in the Fairnesstreatment or theSocietal Beneﬁtstreatment. The reportedp-value fromF-test is for the hypothesis that all Moraltreatments have the same effect. Columns (2), (4), and (6) include the following controls: the amount of self-reported foreign income for the baseline year, the amount of foreign income recorded in the ACA-reports for 2011 and 2012 (scaled in units of one million NOK), gender, age (an indicator variable taking the value one if the taxpayer is more than 60 years old), an indicator variable taking the value one if the taxpayer is a Norwegian citizen, and an indicator variable of socioeconomic status taking the value one if the taxpayer is in the upper 25% of the income and wealth distribution in the baseline year. SES, socioeconomic status.

Robust standard errors are in parentheses: *p<0.1; **p<0.05; ***p<0.01.

(13)

taxpayers do, and this social comparison could also increase compliance by appealing to a desire to con- form to the social norm that tax compliance is socially considered as appropriate behavior (Cullis and Lewis 1997, Wenzel2004, Bartke et al.2017). In this respect, theFairnesstreatment may affect tax compliance through both a focusing effect, by drawing taxpayers’attention to the norm, and an information effect, by making taxpayers aware of how many others follow the norm (Krupka and Weber2009).

As shown in columns (5) and (6) of Table6, where we pool theMoraltreatments, moral suasion on average causes an increase in self-reported foreign income of almost 70% (p0.008, column (6)). The Detection treatment increases average self-reported foreign income by 80% (p0.028, column (6)).

Result 2. Including moral suasion or information that increases the detection probability in the letter from the tax authorities has an economically and statistically highly signiﬁcant effect on the amount of self-reported foreign income.

The strong effect of the moral appeals in our study is likely to reﬂect the fact that we consider a sample of taxpayers who had the opportunity and willingness to evade taxes, as well as the fact that the taxpayers were exposed to the moral appeals close to the deadline for submitting their tax return. Furthermore, context-speciﬁc factors might have made the moral

appeals particularly effective in our experiment. In particular, the social beneﬁt argument and the fairness argument are likely to be more effective in a country where it is generally recognized that taxes ﬁnance important public goods and where the general level of compliance is high (Mascagni2018). The high level of trust in Norway might also have con- tributed to making the moral appeals more effective than what is found in previous studies (Kirchler2007, Kirchler et al.2008).

In panel (b) of Figure1, we show the effect of the letters on the extensive margin. We observe that the base letter signiﬁcantly increases the share of taxpayers who report a positive foreign income for the follow-up year. The moral letter only has a small effect on the extensive margin compared with the base letter, whereas the detection letter causes a large increase in the share reporting a positive amount. Thus, consistent with our theoretical framework, theMoral andDetectiontreatments affect the extensive margin very differently. In Table 7, we report the corresponding regression analysis. In columns (1) and (2), we observe that receiving a letter from the tax authorities increases the share of taxpayers who report a positive foreign income from 11% in the No letter group to 22% in the treatment groups combined, and, as shown in column (2), the effect is robust to the inclusion of the set of background variables (p<0.001, column (2)).

Figure 2. Requests for Deductions of Taxes Paid Abroad

Note. Theﬁgure shows, for the follow-up year, the average amount of requested deductions in Norwegian taxes in NOK (panel (a)), the share that requested deductions (panel (b)), and the share that reported positive amounts of foreign income without requesting any deductions (panel (c)), by treatment.